1 // SPDX-License-Identifier: GPL-2.0
3 * OS Noise Tracer: computes the OS Noise suffered by a running thread.
4 * Timerlat Tracer: measures the wakeup latency of a timer triggered IRQ and thread.
6 * Based on "hwlat_detector" tracer by:
7 * Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com>
8 * Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com>
9 * With feedback from Clark Williams <williams@redhat.com>
11 * And also based on the rtsl tracer presented on:
12 * DE OLIVEIRA, Daniel Bristot, et al. Demystifying the real-time linux
13 * scheduling latency. In: 32nd Euromicro Conference on Real-Time Systems
14 * (ECRTS 2020). Schloss Dagstuhl-Leibniz-Zentrum fur Informatik, 2020.
16 * Copyright (C) 2021 Daniel Bristot de Oliveira, Red Hat, Inc. <bristot@redhat.com>
19 #include <linux/kthread.h>
20 #include <linux/tracefs.h>
21 #include <linux/uaccess.h>
22 #include <linux/cpumask.h>
23 #include <linux/delay.h>
24 #include <linux/sched/clock.h>
25 #include <uapi/linux/sched/types.h>
26 #include <linux/sched.h>
29 #ifdef CONFIG_X86_LOCAL_APIC
30 #include <asm/trace/irq_vectors.h>
31 #undef TRACE_INCLUDE_PATH
32 #undef TRACE_INCLUDE_FILE
33 #endif /* CONFIG_X86_LOCAL_APIC */
35 #include <trace/events/irq.h>
36 #include <trace/events/sched.h>
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/osnoise.h>
44 #define BANNER "osnoise: "
45 #define DEFAULT_SAMPLE_PERIOD 1000000 /* 1s */
46 #define DEFAULT_SAMPLE_RUNTIME 1000000 /* 1s */
48 #define DEFAULT_TIMERLAT_PERIOD 1000 /* 1ms */
49 #define DEFAULT_TIMERLAT_PRIO 95 /* FIFO 95 */
52 * trace_array of the enabled osnoise/timerlat instances.
54 struct osnoise_instance {
55 struct list_head list;
56 struct trace_array *tr;
58 struct list_head osnoise_instances;
60 static bool osnoise_has_registered_instances(void)
62 return !!list_first_or_null_rcu(&osnoise_instances,
63 struct osnoise_instance,
68 * osnoise_instance_registered - check if a tr is already registered
70 static int osnoise_instance_registered(struct trace_array *tr)
72 struct osnoise_instance *inst;
76 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
86 * osnoise_register_instance - register a new trace instance
88 * Register a trace_array *tr in the list of instances running
89 * osnoise/timerlat tracers.
91 static int osnoise_register_instance(struct trace_array *tr)
93 struct osnoise_instance *inst;
96 * register/unregister serialization is provided by trace's
99 lockdep_assert_held(&trace_types_lock);
101 inst = kmalloc(sizeof(*inst), GFP_KERNEL);
105 INIT_LIST_HEAD_RCU(&inst->list);
107 list_add_tail_rcu(&inst->list, &osnoise_instances);
113 * osnoise_unregister_instance - unregister a registered trace instance
115 * Remove the trace_array *tr from the list of instances running
116 * osnoise/timerlat tracers.
118 static void osnoise_unregister_instance(struct trace_array *tr)
120 struct osnoise_instance *inst;
124 * register/unregister serialization is provided by trace's
127 lockdep_assert_held(&trace_types_lock);
129 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
130 if (inst->tr == tr) {
131 list_del_rcu(&inst->list);
161 #define IRQ_CONTEXT 0
162 #define THREAD_CONTEXT 1
164 * sofirq runtime info.
173 * thread runtime info.
182 * Runtime information: this structure saves the runtime information used by
183 * one sampling thread.
185 struct osnoise_variables {
186 struct task_struct *kthread;
191 struct osn_softirq softirq;
192 struct osn_thread thread;
197 * Per-cpu runtime information.
199 DEFINE_PER_CPU(struct osnoise_variables, per_cpu_osnoise_var);
202 * this_cpu_osn_var - Return the per-cpu osnoise_variables on its relative CPU
204 static inline struct osnoise_variables *this_cpu_osn_var(void)
206 return this_cpu_ptr(&per_cpu_osnoise_var);
209 #ifdef CONFIG_TIMERLAT_TRACER
211 * Runtime information for the timer mode.
213 struct timerlat_variables {
214 struct task_struct *kthread;
215 struct hrtimer timer;
222 DEFINE_PER_CPU(struct timerlat_variables, per_cpu_timerlat_var);
225 * this_cpu_tmr_var - Return the per-cpu timerlat_variables on its relative CPU
227 static inline struct timerlat_variables *this_cpu_tmr_var(void)
229 return this_cpu_ptr(&per_cpu_timerlat_var);
233 * tlat_var_reset - Reset the values of the given timerlat_variables
235 static inline void tlat_var_reset(void)
237 struct timerlat_variables *tlat_var;
240 * So far, all the values are initialized as 0, so
241 * zeroing the structure is perfect.
243 for_each_cpu(cpu, cpu_online_mask) {
244 tlat_var = per_cpu_ptr(&per_cpu_timerlat_var, cpu);
245 memset(tlat_var, 0, sizeof(*tlat_var));
248 #else /* CONFIG_TIMERLAT_TRACER */
249 #define tlat_var_reset() do {} while (0)
250 #endif /* CONFIG_TIMERLAT_TRACER */
253 * osn_var_reset - Reset the values of the given osnoise_variables
255 static inline void osn_var_reset(void)
257 struct osnoise_variables *osn_var;
261 * So far, all the values are initialized as 0, so
262 * zeroing the structure is perfect.
264 for_each_cpu(cpu, cpu_online_mask) {
265 osn_var = per_cpu_ptr(&per_cpu_osnoise_var, cpu);
266 memset(osn_var, 0, sizeof(*osn_var));
271 * osn_var_reset_all - Reset the value of all per-cpu osnoise_variables
273 static inline void osn_var_reset_all(void)
280 * Tells NMIs to call back to the osnoise tracer to record timestamps.
282 bool trace_osnoise_callback_enabled;
285 * osnoise sample structure definition. Used to store the statistics of a
288 struct osnoise_sample {
289 u64 runtime; /* runtime */
290 u64 noise; /* noise */
291 u64 max_sample; /* max single noise sample */
292 int hw_count; /* # HW (incl. hypervisor) interference */
293 int nmi_count; /* # NMIs during this sample */
294 int irq_count; /* # IRQs during this sample */
295 int softirq_count; /* # softirqs during this sample */
296 int thread_count; /* # threads during this sample */
299 #ifdef CONFIG_TIMERLAT_TRACER
301 * timerlat sample structure definition. Used to store the statistics of
304 struct timerlat_sample {
305 u64 timer_latency; /* timer_latency */
306 unsigned int seqnum; /* unique sequence */
307 int context; /* timer context */
312 * Protect the interface.
314 struct mutex interface_lock;
319 static struct osnoise_data {
320 u64 sample_period; /* total sampling period */
321 u64 sample_runtime; /* active sampling portion of period */
322 u64 stop_tracing; /* stop trace in the internal operation (loop/irq) */
323 u64 stop_tracing_total; /* stop trace in the final operation (report/thread) */
324 #ifdef CONFIG_TIMERLAT_TRACER
325 u64 timerlat_period; /* timerlat period */
326 u64 print_stack; /* print IRQ stack if total > */
327 int timerlat_tracer; /* timerlat tracer */
329 bool tainted; /* infor users and developers about a problem */
331 .sample_period = DEFAULT_SAMPLE_PERIOD,
332 .sample_runtime = DEFAULT_SAMPLE_RUNTIME,
334 .stop_tracing_total = 0,
335 #ifdef CONFIG_TIMERLAT_TRACER
337 .timerlat_period = DEFAULT_TIMERLAT_PERIOD,
338 .timerlat_tracer = 0,
342 #ifdef CONFIG_TIMERLAT_TRACER
343 static inline bool timerlat_enabled(void)
345 return osnoise_data.timerlat_tracer;
348 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
350 struct timerlat_variables *tlat_var = this_cpu_tmr_var();
352 * If the timerlat is enabled, but the irq handler did
353 * not run yet enabling timerlat_tracer, do not trace.
355 if (!tlat_var->tracing_thread) {
356 osn_var->softirq.arrival_time = 0;
357 osn_var->softirq.delta_start = 0;
363 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
365 struct timerlat_variables *tlat_var = this_cpu_tmr_var();
367 * If the timerlat is enabled, but the irq handler did
368 * not run yet enabling timerlat_tracer, do not trace.
370 if (!tlat_var->tracing_thread) {
371 osn_var->thread.delta_start = 0;
372 osn_var->thread.arrival_time = 0;
377 #else /* CONFIG_TIMERLAT_TRACER */
378 static inline bool timerlat_enabled(void)
383 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
387 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
393 #ifdef CONFIG_PREEMPT_RT
395 * Print the osnoise header info.
397 static void print_osnoise_headers(struct seq_file *s)
399 if (osnoise_data.tainted)
400 seq_puts(s, "# osnoise is tainted!\n");
402 seq_puts(s, "# _-------=> irqs-off\n");
403 seq_puts(s, "# / _------=> need-resched\n");
404 seq_puts(s, "# | / _-----=> need-resched-lazy\n");
405 seq_puts(s, "# || / _----=> hardirq/softirq\n");
406 seq_puts(s, "# ||| / _---=> preempt-depth\n");
407 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n");
408 seq_puts(s, "# ||||| / _-=> migrate-disable\n");
410 seq_puts(s, "# |||||| / ");
411 seq_puts(s, " MAX\n");
413 seq_puts(s, "# ||||| / ");
414 seq_puts(s, " SINGLE Interference counters:\n");
416 seq_puts(s, "# ||||||| RUNTIME ");
417 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n");
419 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP IN US ");
420 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n");
422 seq_puts(s, "# | | | ||||||| | | ");
423 seq_puts(s, " | | | | | | | |\n");
425 #else /* CONFIG_PREEMPT_RT */
426 static void print_osnoise_headers(struct seq_file *s)
428 if (osnoise_data.tainted)
429 seq_puts(s, "# osnoise is tainted!\n");
431 seq_puts(s, "# _-----=> irqs-off\n");
432 seq_puts(s, "# / _----=> need-resched\n");
433 seq_puts(s, "# | / _---=> hardirq/softirq\n");
434 seq_puts(s, "# || / _--=> preempt-depth\n");
435 seq_puts(s, "# ||| / _-=> migrate-disable ");
436 seq_puts(s, " MAX\n");
437 seq_puts(s, "# |||| / delay ");
438 seq_puts(s, " SINGLE Interference counters:\n");
440 seq_puts(s, "# ||||| RUNTIME ");
441 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n");
443 seq_puts(s, "# TASK-PID CPU# ||||| TIMESTAMP IN US ");
444 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n");
446 seq_puts(s, "# | | | ||||| | | ");
447 seq_puts(s, " | | | | | | | |\n");
449 #endif /* CONFIG_PREEMPT_RT */
452 * osnoise_taint - report an osnoise error.
454 #define osnoise_taint(msg) ({ \
455 struct osnoise_instance *inst; \
456 struct trace_buffer *buffer; \
459 list_for_each_entry_rcu(inst, &osnoise_instances, list) { \
460 buffer = inst->tr->array_buffer.buffer; \
461 trace_array_printk_buf(buffer, _THIS_IP_, msg); \
464 osnoise_data.tainted = true; \
468 * Record an osnoise_sample into the tracer buffer.
471 __trace_osnoise_sample(struct osnoise_sample *sample, struct trace_buffer *buffer)
473 struct trace_event_call *call = &event_osnoise;
474 struct ring_buffer_event *event;
475 struct osnoise_entry *entry;
477 event = trace_buffer_lock_reserve(buffer, TRACE_OSNOISE, sizeof(*entry),
481 entry = ring_buffer_event_data(event);
482 entry->runtime = sample->runtime;
483 entry->noise = sample->noise;
484 entry->max_sample = sample->max_sample;
485 entry->hw_count = sample->hw_count;
486 entry->nmi_count = sample->nmi_count;
487 entry->irq_count = sample->irq_count;
488 entry->softirq_count = sample->softirq_count;
489 entry->thread_count = sample->thread_count;
491 if (!call_filter_check_discard(call, entry, buffer, event))
492 trace_buffer_unlock_commit_nostack(buffer, event);
496 * Record an osnoise_sample on all osnoise instances.
498 static void trace_osnoise_sample(struct osnoise_sample *sample)
500 struct osnoise_instance *inst;
501 struct trace_buffer *buffer;
504 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
505 buffer = inst->tr->array_buffer.buffer;
506 __trace_osnoise_sample(sample, buffer);
511 #ifdef CONFIG_TIMERLAT_TRACER
513 * Print the timerlat header info.
515 #ifdef CONFIG_PREEMPT_RT
516 static void print_timerlat_headers(struct seq_file *s)
518 seq_puts(s, "# _-------=> irqs-off\n");
519 seq_puts(s, "# / _------=> need-resched\n");
520 seq_puts(s, "# | / _-----=> need-resched-lazy\n");
521 seq_puts(s, "# || / _----=> hardirq/softirq\n");
522 seq_puts(s, "# ||| / _---=> preempt-depth\n");
523 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n");
524 seq_puts(s, "# ||||| / _-=> migrate-disable\n");
525 seq_puts(s, "# |||||| /\n");
526 seq_puts(s, "# ||||||| ACTIVATION\n");
527 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP ID ");
528 seq_puts(s, " CONTEXT LATENCY\n");
529 seq_puts(s, "# | | | ||||||| | | ");
530 seq_puts(s, " | |\n");
532 #else /* CONFIG_PREEMPT_RT */
533 static void print_timerlat_headers(struct seq_file *s)
535 seq_puts(s, "# _-----=> irqs-off\n");
536 seq_puts(s, "# / _----=> need-resched\n");
537 seq_puts(s, "# | / _---=> hardirq/softirq\n");
538 seq_puts(s, "# || / _--=> preempt-depth\n");
539 seq_puts(s, "# ||| / _-=> migrate-disable\n");
540 seq_puts(s, "# |||| / delay\n");
541 seq_puts(s, "# ||||| ACTIVATION\n");
542 seq_puts(s, "# TASK-PID CPU# ||||| TIMESTAMP ID ");
543 seq_puts(s, " CONTEXT LATENCY\n");
544 seq_puts(s, "# | | | ||||| | | ");
545 seq_puts(s, " | |\n");
547 #endif /* CONFIG_PREEMPT_RT */
550 __trace_timerlat_sample(struct timerlat_sample *sample, struct trace_buffer *buffer)
552 struct trace_event_call *call = &event_osnoise;
553 struct ring_buffer_event *event;
554 struct timerlat_entry *entry;
556 event = trace_buffer_lock_reserve(buffer, TRACE_TIMERLAT, sizeof(*entry),
560 entry = ring_buffer_event_data(event);
561 entry->seqnum = sample->seqnum;
562 entry->context = sample->context;
563 entry->timer_latency = sample->timer_latency;
565 if (!call_filter_check_discard(call, entry, buffer, event))
566 trace_buffer_unlock_commit_nostack(buffer, event);
570 * Record an timerlat_sample into the tracer buffer.
572 static void trace_timerlat_sample(struct timerlat_sample *sample)
574 struct osnoise_instance *inst;
575 struct trace_buffer *buffer;
578 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
579 buffer = inst->tr->array_buffer.buffer;
580 __trace_timerlat_sample(sample, buffer);
585 #ifdef CONFIG_STACKTRACE
587 #define MAX_CALLS 256
590 * Stack trace will take place only at IRQ level, so, no need
591 * to control nesting here.
596 unsigned long calls[MAX_CALLS];
599 static DEFINE_PER_CPU(struct trace_stack, trace_stack);
602 * timerlat_save_stack - save a stack trace without printing
604 * Save the current stack trace without printing. The
605 * stack will be printed later, after the end of the measurement.
607 static void timerlat_save_stack(int skip)
609 unsigned int size, nr_entries;
610 struct trace_stack *fstack;
612 fstack = this_cpu_ptr(&trace_stack);
614 size = ARRAY_SIZE(fstack->calls);
616 nr_entries = stack_trace_save(fstack->calls, size, skip);
618 fstack->stack_size = nr_entries * sizeof(unsigned long);
619 fstack->nr_entries = nr_entries;
626 __timerlat_dump_stack(struct trace_buffer *buffer, struct trace_stack *fstack, unsigned int size)
628 struct trace_event_call *call = &event_osnoise;
629 struct ring_buffer_event *event;
630 struct stack_entry *entry;
632 event = trace_buffer_lock_reserve(buffer, TRACE_STACK, sizeof(*entry) + size,
637 entry = ring_buffer_event_data(event);
639 memcpy(&entry->caller, fstack->calls, size);
640 entry->size = fstack->nr_entries;
642 if (!call_filter_check_discard(call, entry, buffer, event))
643 trace_buffer_unlock_commit_nostack(buffer, event);
647 * timerlat_dump_stack - dump a stack trace previously saved
649 static void timerlat_dump_stack(u64 latency)
651 struct osnoise_instance *inst;
652 struct trace_buffer *buffer;
653 struct trace_stack *fstack;
657 * trace only if latency > print_stack config, if enabled.
659 if (!osnoise_data.print_stack || osnoise_data.print_stack > latency)
662 preempt_disable_notrace();
663 fstack = this_cpu_ptr(&trace_stack);
664 size = fstack->stack_size;
667 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
668 buffer = inst->tr->array_buffer.buffer;
669 __timerlat_dump_stack(buffer, fstack, size);
673 preempt_enable_notrace();
675 #else /* CONFIG_STACKTRACE */
676 #define timerlat_dump_stack(u64 latency) do {} while (0)
677 #define timerlat_save_stack(a) do {} while (0)
678 #endif /* CONFIG_STACKTRACE */
679 #endif /* CONFIG_TIMERLAT_TRACER */
682 * Macros to encapsulate the time capturing infrastructure.
684 #define time_get() trace_clock_local()
685 #define time_to_us(x) div_u64(x, 1000)
686 #define time_sub(a, b) ((a) - (b))
689 * cond_move_irq_delta_start - Forward the delta_start of a running IRQ
691 * If an IRQ is preempted by an NMI, its delta_start is pushed forward
692 * to discount the NMI interference.
694 * See get_int_safe_duration().
697 cond_move_irq_delta_start(struct osnoise_variables *osn_var, u64 duration)
699 if (osn_var->irq.delta_start)
700 osn_var->irq.delta_start += duration;
703 #ifndef CONFIG_PREEMPT_RT
705 * cond_move_softirq_delta_start - Forward the delta_start of a running softirq.
707 * If a softirq is preempted by an IRQ or NMI, its delta_start is pushed
708 * forward to discount the interference.
710 * See get_int_safe_duration().
713 cond_move_softirq_delta_start(struct osnoise_variables *osn_var, u64 duration)
715 if (osn_var->softirq.delta_start)
716 osn_var->softirq.delta_start += duration;
718 #else /* CONFIG_PREEMPT_RT */
719 #define cond_move_softirq_delta_start(osn_var, duration) do {} while (0)
723 * cond_move_thread_delta_start - Forward the delta_start of a running thread
725 * If a noisy thread is preempted by an softirq, IRQ or NMI, its delta_start
726 * is pushed forward to discount the interference.
728 * See get_int_safe_duration().
731 cond_move_thread_delta_start(struct osnoise_variables *osn_var, u64 duration)
733 if (osn_var->thread.delta_start)
734 osn_var->thread.delta_start += duration;
738 * get_int_safe_duration - Get the duration of a window
740 * The irq, softirq and thread varaibles need to have its duration without
741 * the interference from higher priority interrupts. Instead of keeping a
742 * variable to discount the interrupt interference from these variables, the
743 * starting time of these variables are pushed forward with the interrupt's
744 * duration. In this way, a single variable is used to:
746 * - Know if a given window is being measured.
747 * - Account its duration.
748 * - Discount the interference.
750 * To avoid getting inconsistent values, e.g.,:
754 * delta_start -= int duration;
756 * duration = now - delta_start;
758 * result: negative duration if the variable duration before the
759 * interrupt was smaller than the interrupt execution.
761 * A counter of interrupts is used. If the counter increased, try
762 * to capture an interference safe duration.
765 get_int_safe_duration(struct osnoise_variables *osn_var, u64 *delta_start)
767 u64 int_counter, now;
771 int_counter = local_read(&osn_var->int_counter);
772 /* synchronize with interrupts */
776 duration = (now - *delta_start);
778 /* synchronize with interrupts */
780 } while (int_counter != local_read(&osn_var->int_counter));
783 * This is an evidence of race conditions that cause
784 * a value to be "discounted" too much.
787 osnoise_taint("Negative duration!\n");
796 * set_int_safe_time - Save the current time on *time, aware of interference
798 * Get the time, taking into consideration a possible interference from
799 * higher priority interrupts.
801 * See get_int_safe_duration() for an explanation.
804 set_int_safe_time(struct osnoise_variables *osn_var, u64 *time)
809 int_counter = local_read(&osn_var->int_counter);
810 /* synchronize with interrupts */
815 /* synchronize with interrupts */
817 } while (int_counter != local_read(&osn_var->int_counter));
822 #ifdef CONFIG_TIMERLAT_TRACER
824 * copy_int_safe_time - Copy *src into *desc aware of interference
827 copy_int_safe_time(struct osnoise_variables *osn_var, u64 *dst, u64 *src)
832 int_counter = local_read(&osn_var->int_counter);
833 /* synchronize with interrupts */
838 /* synchronize with interrupts */
840 } while (int_counter != local_read(&osn_var->int_counter));
844 #endif /* CONFIG_TIMERLAT_TRACER */
847 * trace_osnoise_callback - NMI entry/exit callback
849 * This function is called at the entry and exit NMI code. The bool enter
850 * distinguishes between either case. This function is used to note a NMI
851 * occurrence, compute the noise caused by the NMI, and to remove the noise
852 * it is potentially causing on other interference variables.
854 void trace_osnoise_callback(bool enter)
856 struct osnoise_variables *osn_var = this_cpu_osn_var();
859 if (!osn_var->sampling)
863 * Currently trace_clock_local() calls sched_clock() and the
864 * generic version is not NMI safe.
866 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
868 osn_var->nmi.delta_start = time_get();
869 local_inc(&osn_var->int_counter);
871 duration = time_get() - osn_var->nmi.delta_start;
873 trace_nmi_noise(osn_var->nmi.delta_start, duration);
875 cond_move_irq_delta_start(osn_var, duration);
876 cond_move_softirq_delta_start(osn_var, duration);
877 cond_move_thread_delta_start(osn_var, duration);
882 osn_var->nmi.count++;
886 * osnoise_trace_irq_entry - Note the starting of an IRQ
888 * Save the starting time of an IRQ. As IRQs are non-preemptive to other IRQs,
889 * it is safe to use a single variable (ons_var->irq) to save the statistics.
890 * The arrival_time is used to report... the arrival time. The delta_start
891 * is used to compute the duration at the IRQ exit handler. See
892 * cond_move_irq_delta_start().
894 void osnoise_trace_irq_entry(int id)
896 struct osnoise_variables *osn_var = this_cpu_osn_var();
898 if (!osn_var->sampling)
901 * This value will be used in the report, but not to compute
902 * the execution time, so it is safe to get it unsafe.
904 osn_var->irq.arrival_time = time_get();
905 set_int_safe_time(osn_var, &osn_var->irq.delta_start);
906 osn_var->irq.count++;
908 local_inc(&osn_var->int_counter);
912 * osnoise_irq_exit - Note the end of an IRQ, sava data and trace
914 * Computes the duration of the IRQ noise, and trace it. Also discounts the
915 * interference from other sources of noise could be currently being accounted.
917 void osnoise_trace_irq_exit(int id, const char *desc)
919 struct osnoise_variables *osn_var = this_cpu_osn_var();
922 if (!osn_var->sampling)
925 duration = get_int_safe_duration(osn_var, &osn_var->irq.delta_start);
926 trace_irq_noise(id, desc, osn_var->irq.arrival_time, duration);
927 osn_var->irq.arrival_time = 0;
928 cond_move_softirq_delta_start(osn_var, duration);
929 cond_move_thread_delta_start(osn_var, duration);
933 * trace_irqentry_callback - Callback to the irq:irq_entry traceevent
935 * Used to note the starting of an IRQ occurece.
937 static void trace_irqentry_callback(void *data, int irq,
938 struct irqaction *action)
940 osnoise_trace_irq_entry(irq);
944 * trace_irqexit_callback - Callback to the irq:irq_exit traceevent
946 * Used to note the end of an IRQ occurece.
948 static void trace_irqexit_callback(void *data, int irq,
949 struct irqaction *action, int ret)
951 osnoise_trace_irq_exit(irq, action->name);
955 * arch specific register function.
957 int __weak osnoise_arch_register(void)
963 * arch specific unregister function.
965 void __weak osnoise_arch_unregister(void)
971 * hook_irq_events - Hook IRQ handling events
973 * This function hooks the IRQ related callbacks to the respective trace
976 static int hook_irq_events(void)
980 ret = register_trace_irq_handler_entry(trace_irqentry_callback, NULL);
984 ret = register_trace_irq_handler_exit(trace_irqexit_callback, NULL);
986 goto out_unregister_entry;
988 ret = osnoise_arch_register();
995 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
996 out_unregister_entry:
997 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1003 * unhook_irq_events - Unhook IRQ handling events
1005 * This function unhooks the IRQ related callbacks to the respective trace
1008 static void unhook_irq_events(void)
1010 osnoise_arch_unregister();
1011 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
1012 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1015 #ifndef CONFIG_PREEMPT_RT
1017 * trace_softirq_entry_callback - Note the starting of a softirq
1019 * Save the starting time of a softirq. As softirqs are non-preemptive to
1020 * other softirqs, it is safe to use a single variable (ons_var->softirq)
1021 * to save the statistics. The arrival_time is used to report... the
1022 * arrival time. The delta_start is used to compute the duration at the
1023 * softirq exit handler. See cond_move_softirq_delta_start().
1025 static void trace_softirq_entry_callback(void *data, unsigned int vec_nr)
1027 struct osnoise_variables *osn_var = this_cpu_osn_var();
1029 if (!osn_var->sampling)
1032 * This value will be used in the report, but not to compute
1033 * the execution time, so it is safe to get it unsafe.
1035 osn_var->softirq.arrival_time = time_get();
1036 set_int_safe_time(osn_var, &osn_var->softirq.delta_start);
1037 osn_var->softirq.count++;
1039 local_inc(&osn_var->int_counter);
1043 * trace_softirq_exit_callback - Note the end of an softirq
1045 * Computes the duration of the softirq noise, and trace it. Also discounts the
1046 * interference from other sources of noise could be currently being accounted.
1048 static void trace_softirq_exit_callback(void *data, unsigned int vec_nr)
1050 struct osnoise_variables *osn_var = this_cpu_osn_var();
1053 if (!osn_var->sampling)
1056 if (unlikely(timerlat_enabled()))
1057 if (!timerlat_softirq_exit(osn_var))
1060 duration = get_int_safe_duration(osn_var, &osn_var->softirq.delta_start);
1061 trace_softirq_noise(vec_nr, osn_var->softirq.arrival_time, duration);
1062 cond_move_thread_delta_start(osn_var, duration);
1063 osn_var->softirq.arrival_time = 0;
1067 * hook_softirq_events - Hook softirq handling events
1069 * This function hooks the softirq related callbacks to the respective trace
1072 static int hook_softirq_events(void)
1076 ret = register_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1080 ret = register_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1082 goto out_unreg_entry;
1087 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1093 * unhook_softirq_events - Unhook softirq handling events
1095 * This function hooks the softirq related callbacks to the respective trace
1098 static void unhook_softirq_events(void)
1100 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1101 unregister_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1103 #else /* CONFIG_PREEMPT_RT */
1105 * softirq are threads on the PREEMPT_RT mode.
1107 static int hook_softirq_events(void)
1111 static void unhook_softirq_events(void)
1117 * thread_entry - Record the starting of a thread noise window
1119 * It saves the context switch time for a noisy thread, and increments
1120 * the interference counters.
1123 thread_entry(struct osnoise_variables *osn_var, struct task_struct *t)
1125 if (!osn_var->sampling)
1128 * The arrival time will be used in the report, but not to compute
1129 * the execution time, so it is safe to get it unsafe.
1131 osn_var->thread.arrival_time = time_get();
1133 set_int_safe_time(osn_var, &osn_var->thread.delta_start);
1135 osn_var->thread.count++;
1136 local_inc(&osn_var->int_counter);
1140 * thread_exit - Report the end of a thread noise window
1142 * It computes the total noise from a thread, tracing if needed.
1145 thread_exit(struct osnoise_variables *osn_var, struct task_struct *t)
1149 if (!osn_var->sampling)
1152 if (unlikely(timerlat_enabled()))
1153 if (!timerlat_thread_exit(osn_var))
1156 duration = get_int_safe_duration(osn_var, &osn_var->thread.delta_start);
1158 trace_thread_noise(t, osn_var->thread.arrival_time, duration);
1160 osn_var->thread.arrival_time = 0;
1164 * trace_sched_switch - sched:sched_switch trace event handler
1166 * This function is hooked to the sched:sched_switch trace event, and it is
1167 * used to record the beginning and to report the end of a thread noise window.
1170 trace_sched_switch_callback(void *data, bool preempt, struct task_struct *p,
1171 struct task_struct *n)
1173 struct osnoise_variables *osn_var = this_cpu_osn_var();
1175 if (p->pid != osn_var->pid)
1176 thread_exit(osn_var, p);
1178 if (n->pid != osn_var->pid)
1179 thread_entry(osn_var, n);
1183 * hook_thread_events - Hook the insturmentation for thread noise
1185 * Hook the osnoise tracer callbacks to handle the noise from other
1186 * threads on the necessary kernel events.
1188 static int hook_thread_events(void)
1192 ret = register_trace_sched_switch(trace_sched_switch_callback, NULL);
1200 * unhook_thread_events - *nhook the insturmentation for thread noise
1202 * Unook the osnoise tracer callbacks to handle the noise from other
1203 * threads on the necessary kernel events.
1205 static void unhook_thread_events(void)
1207 unregister_trace_sched_switch(trace_sched_switch_callback, NULL);
1211 * save_osn_sample_stats - Save the osnoise_sample statistics
1213 * Save the osnoise_sample statistics before the sampling phase. These
1214 * values will be used later to compute the diff betwneen the statistics
1215 * before and after the osnoise sampling.
1218 save_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1220 s->nmi_count = osn_var->nmi.count;
1221 s->irq_count = osn_var->irq.count;
1222 s->softirq_count = osn_var->softirq.count;
1223 s->thread_count = osn_var->thread.count;
1227 * diff_osn_sample_stats - Compute the osnoise_sample statistics
1229 * After a sample period, compute the difference on the osnoise_sample
1230 * statistics. The struct osnoise_sample *s contains the statistics saved via
1231 * save_osn_sample_stats() before the osnoise sampling.
1234 diff_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1236 s->nmi_count = osn_var->nmi.count - s->nmi_count;
1237 s->irq_count = osn_var->irq.count - s->irq_count;
1238 s->softirq_count = osn_var->softirq.count - s->softirq_count;
1239 s->thread_count = osn_var->thread.count - s->thread_count;
1243 * osnoise_stop_tracing - Stop tracing and the tracer.
1245 static __always_inline void osnoise_stop_tracing(void)
1247 struct osnoise_instance *inst;
1248 struct trace_array *tr;
1251 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1253 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_,
1254 "stop tracing hit on cpu %d\n", smp_processor_id());
1256 tracer_tracing_off(tr);
1262 * notify_new_max_latency - Notify a new max latency via fsnotify interface.
1264 static void notify_new_max_latency(u64 latency)
1266 struct osnoise_instance *inst;
1267 struct trace_array *tr;
1270 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1272 if (tr->max_latency < latency) {
1273 tr->max_latency = latency;
1274 latency_fsnotify(tr);
1281 * run_osnoise - Sample the time and look for osnoise
1283 * Used to capture the time, looking for potential osnoise latency repeatedly.
1284 * Different from hwlat_detector, it is called with preemption and interrupts
1285 * enabled. This allows irqs, softirqs and threads to run, interfering on the
1286 * osnoise sampling thread, as they would do with a regular thread.
1288 static int run_osnoise(void)
1290 struct osnoise_variables *osn_var = this_cpu_osn_var();
1291 u64 start, sample, last_sample;
1292 u64 last_int_count, int_count;
1293 s64 noise = 0, max_noise = 0;
1294 s64 total, last_total = 0;
1295 struct osnoise_sample s;
1296 unsigned int threshold;
1297 u64 runtime, stop_in;
1303 * Considers the current thread as the workload.
1305 osn_var->pid = current->pid;
1308 * Save the current stats for the diff
1310 save_osn_sample_stats(osn_var, &s);
1313 * if threshold is 0, use the default value of 5 us.
1315 threshold = tracing_thresh ? : 5000;
1318 * Make sure NMIs see sampling first
1320 osn_var->sampling = true;
1324 * Transform the *_us config to nanoseconds to avoid the
1325 * division on the main loop.
1327 runtime = osnoise_data.sample_runtime * NSEC_PER_USEC;
1328 stop_in = osnoise_data.stop_tracing * NSEC_PER_USEC;
1338 last_int_count = set_int_safe_time(osn_var, &last_sample);
1344 int_count = set_int_safe_time(osn_var, &sample);
1346 noise = time_sub(sample, last_sample);
1349 * This shouldn't happen.
1352 osnoise_taint("negative noise!");
1359 total = time_sub(sample, start);
1362 * Check for possible overflows.
1364 if (total < last_total) {
1365 osnoise_taint("total overflow!");
1371 if (noise >= threshold) {
1372 int interference = int_count - last_int_count;
1374 if (noise > max_noise)
1382 trace_sample_threshold(last_sample, noise, interference);
1384 if (osnoise_data.stop_tracing)
1385 if (noise > stop_in)
1386 osnoise_stop_tracing();
1390 * For the non-preemptive kernel config: let threads runs, if
1395 last_sample = sample;
1396 last_int_count = int_count;
1398 } while (total < runtime && !kthread_should_stop());
1401 * Finish the above in the view for interrupts.
1405 osn_var->sampling = false;
1408 * Make sure sampling data is no longer updated.
1415 s.noise = time_to_us(sum_noise);
1416 s.runtime = time_to_us(total);
1417 s.max_sample = time_to_us(max_noise);
1418 s.hw_count = hw_count;
1420 /* Save interference stats info */
1421 diff_osn_sample_stats(osn_var, &s);
1423 trace_osnoise_sample(&s);
1425 notify_new_max_latency(max_noise);
1427 if (osnoise_data.stop_tracing_total)
1428 if (s.noise > osnoise_data.stop_tracing_total)
1429 osnoise_stop_tracing();
1436 static struct cpumask osnoise_cpumask;
1437 static struct cpumask save_cpumask;
1440 * osnoise_main - The osnoise detection kernel thread
1442 * Calls run_osnoise() function to measure the osnoise for the configured runtime,
1445 static int osnoise_main(void *data)
1449 while (!kthread_should_stop()) {
1453 mutex_lock(&interface_lock);
1454 interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
1455 mutex_unlock(&interface_lock);
1457 do_div(interval, USEC_PER_MSEC);
1460 * differently from hwlat_detector, the osnoise tracer can run
1461 * without a pause because preemption is on.
1464 /* Let synchronize_rcu_tasks() make progress */
1465 cond_resched_tasks_rcu_qs();
1469 if (msleep_interruptible(interval))
1476 #ifdef CONFIG_TIMERLAT_TRACER
1478 * timerlat_irq - hrtimer handler for timerlat.
1480 static enum hrtimer_restart timerlat_irq(struct hrtimer *timer)
1482 struct osnoise_variables *osn_var = this_cpu_osn_var();
1483 struct timerlat_variables *tlat;
1484 struct timerlat_sample s;
1489 * I am not sure if the timer was armed for this CPU. So, get
1490 * the timerlat struct from the timer itself, not from this
1493 tlat = container_of(timer, struct timerlat_variables, timer);
1495 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1498 * Enable the osnoise: events for thread an softirq.
1500 tlat->tracing_thread = true;
1502 osn_var->thread.arrival_time = time_get();
1505 * A hardirq is running: the timer IRQ. It is for sure preempting
1506 * a thread, and potentially preempting a softirq.
1508 * At this point, it is not interesting to know the duration of the
1509 * preempted thread (and maybe softirq), but how much time they will
1510 * delay the beginning of the execution of the timer thread.
1512 * To get the correct (net) delay added by the softirq, its delta_start
1513 * is set as the IRQ one. In this way, at the return of the IRQ, the delta
1514 * start of the sofitrq will be zeroed, accounting then only the time
1517 * The thread follows the same principle. However, if a softirq is
1518 * running, the thread needs to receive the softirq delta_start. The
1519 * reason being is that the softirq will be the last to be unfolded,
1520 * resseting the thread delay to zero.
1522 * The PREEMPT_RT is a special case, though. As softirqs run as threads
1523 * on RT, moving the thread is enough.
1525 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && osn_var->softirq.delta_start) {
1526 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1527 &osn_var->softirq.delta_start);
1529 copy_int_safe_time(osn_var, &osn_var->softirq.delta_start,
1530 &osn_var->irq.delta_start);
1532 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1533 &osn_var->irq.delta_start);
1537 * Compute the current time with the expected time.
1539 diff = now - tlat->abs_period;
1542 s.seqnum = tlat->count;
1543 s.timer_latency = diff;
1544 s.context = IRQ_CONTEXT;
1546 trace_timerlat_sample(&s);
1548 notify_new_max_latency(diff);
1550 if (osnoise_data.stop_tracing)
1551 if (time_to_us(diff) >= osnoise_data.stop_tracing)
1552 osnoise_stop_tracing();
1554 wake_up_process(tlat->kthread);
1556 if (osnoise_data.print_stack)
1557 timerlat_save_stack(0);
1559 return HRTIMER_NORESTART;
1563 * wait_next_period - Wait for the next period for timerlat
1565 static int wait_next_period(struct timerlat_variables *tlat)
1567 ktime_t next_abs_period, now;
1568 u64 rel_period = osnoise_data.timerlat_period * 1000;
1570 now = hrtimer_cb_get_time(&tlat->timer);
1571 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1574 * Save the next abs_period.
1576 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1579 * If the new abs_period is in the past, skip the activation.
1581 while (ktime_compare(now, next_abs_period) > 0) {
1582 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1583 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1586 set_current_state(TASK_INTERRUPTIBLE);
1588 hrtimer_start(&tlat->timer, next_abs_period, HRTIMER_MODE_ABS_PINNED_HARD);
1594 * timerlat_main- Timerlat main
1596 static int timerlat_main(void *data)
1598 struct osnoise_variables *osn_var = this_cpu_osn_var();
1599 struct timerlat_variables *tlat = this_cpu_tmr_var();
1600 struct timerlat_sample s;
1601 struct sched_param sp;
1605 * Make the thread RT, that is how cyclictest is usually used.
1607 sp.sched_priority = DEFAULT_TIMERLAT_PRIO;
1608 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1611 tlat->tracing_thread = false;
1613 hrtimer_init(&tlat->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD);
1614 tlat->timer.function = timerlat_irq;
1615 tlat->kthread = current;
1616 osn_var->pid = current->pid;
1618 * Anotate the arrival time.
1620 tlat->abs_period = hrtimer_cb_get_time(&tlat->timer);
1622 wait_next_period(tlat);
1624 osn_var->sampling = 1;
1626 while (!kthread_should_stop()) {
1627 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1628 diff = now - tlat->abs_period;
1630 s.seqnum = tlat->count;
1631 s.timer_latency = diff;
1632 s.context = THREAD_CONTEXT;
1634 trace_timerlat_sample(&s);
1636 timerlat_dump_stack(time_to_us(diff));
1638 tlat->tracing_thread = false;
1639 if (osnoise_data.stop_tracing_total)
1640 if (time_to_us(diff) >= osnoise_data.stop_tracing_total)
1641 osnoise_stop_tracing();
1643 wait_next_period(tlat);
1646 hrtimer_cancel(&tlat->timer);
1649 #else /* CONFIG_TIMERLAT_TRACER */
1650 static int timerlat_main(void *data)
1654 #endif /* CONFIG_TIMERLAT_TRACER */
1657 * stop_kthread - stop a workload thread
1659 static void stop_kthread(unsigned int cpu)
1661 struct task_struct *kthread;
1663 kthread = per_cpu(per_cpu_osnoise_var, cpu).kthread;
1665 kthread_stop(kthread);
1666 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1670 * stop_per_cpu_kthread - Stop per-cpu threads
1672 * Stop the osnoise sampling htread. Use this on unload and at system
1675 static void stop_per_cpu_kthreads(void)
1681 for_each_online_cpu(cpu)
1688 * start_kthread - Start a workload tread
1690 static int start_kthread(unsigned int cpu)
1692 struct task_struct *kthread;
1693 void *main = osnoise_main;
1696 if (timerlat_enabled()) {
1697 snprintf(comm, 24, "timerlat/%d", cpu);
1698 main = timerlat_main;
1700 snprintf(comm, 24, "osnoise/%d", cpu);
1703 kthread = kthread_create_on_cpu(main, NULL, cpu, comm);
1705 if (IS_ERR(kthread)) {
1706 pr_err(BANNER "could not start sampling thread\n");
1707 stop_per_cpu_kthreads();
1711 per_cpu(per_cpu_osnoise_var, cpu).kthread = kthread;
1712 wake_up_process(kthread);
1718 * start_per_cpu_kthread - Kick off per-cpu osnoise sampling kthreads
1720 * This starts the kernel thread that will look for osnoise on many
1723 static int start_per_cpu_kthreads(void)
1725 struct cpumask *current_mask = &save_cpumask;
1731 * Run only on online CPUs in which osnoise is allowed to run.
1733 cpumask_and(current_mask, cpu_online_mask, &osnoise_cpumask);
1735 for_each_possible_cpu(cpu)
1736 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1738 for_each_cpu(cpu, current_mask) {
1739 retval = start_kthread(cpu);
1741 stop_per_cpu_kthreads();
1751 #ifdef CONFIG_HOTPLUG_CPU
1752 static void osnoise_hotplug_workfn(struct work_struct *dummy)
1754 unsigned int cpu = smp_processor_id();
1756 mutex_lock(&trace_types_lock);
1758 if (!osnoise_has_registered_instances())
1759 goto out_unlock_trace;
1761 mutex_lock(&interface_lock);
1764 if (!cpumask_test_cpu(cpu, &osnoise_cpumask))
1771 mutex_unlock(&interface_lock);
1773 mutex_unlock(&trace_types_lock);
1776 static DECLARE_WORK(osnoise_hotplug_work, osnoise_hotplug_workfn);
1779 * osnoise_cpu_init - CPU hotplug online callback function
1781 static int osnoise_cpu_init(unsigned int cpu)
1783 schedule_work_on(cpu, &osnoise_hotplug_work);
1788 * osnoise_cpu_die - CPU hotplug offline callback function
1790 static int osnoise_cpu_die(unsigned int cpu)
1796 static void osnoise_init_hotplug_support(void)
1800 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/osnoise:online",
1801 osnoise_cpu_init, osnoise_cpu_die);
1803 pr_warn(BANNER "Error to init cpu hotplug support\n");
1807 #else /* CONFIG_HOTPLUG_CPU */
1808 static void osnoise_init_hotplug_support(void)
1812 #endif /* CONFIG_HOTPLUG_CPU */
1815 * osnoise_cpus_read - Read function for reading the "cpus" file
1816 * @filp: The active open file structure
1817 * @ubuf: The userspace provided buffer to read value into
1818 * @cnt: The maximum number of bytes to read
1819 * @ppos: The current "file" position
1821 * Prints the "cpus" output into the user-provided buffer.
1824 osnoise_cpus_read(struct file *filp, char __user *ubuf, size_t count,
1830 mutex_lock(&interface_lock);
1832 len = snprintf(NULL, 0, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)) + 1;
1833 mask_str = kmalloc(len, GFP_KERNEL);
1839 len = snprintf(mask_str, len, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask));
1845 count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len);
1850 mutex_unlock(&interface_lock);
1856 * osnoise_cpus_write - Write function for "cpus" entry
1857 * @filp: The active open file structure
1858 * @ubuf: The user buffer that contains the value to write
1859 * @cnt: The maximum number of bytes to write to "file"
1860 * @ppos: The current position in @file
1862 * This function provides a write implementation for the "cpus"
1863 * interface to the osnoise trace. By default, it lists all CPUs,
1864 * in this way, allowing osnoise threads to run on any online CPU
1865 * of the system. It serves to restrict the execution of osnoise to the
1866 * set of CPUs writing via this interface. Why not use "tracing_cpumask"?
1867 * Because the user might be interested in tracing what is running on
1868 * other CPUs. For instance, one might run osnoise in one HT CPU
1869 * while observing what is running on the sibling HT CPU.
1872 osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count,
1875 cpumask_var_t osnoise_cpumask_new;
1882 if (copy_from_user(buf, ubuf, count))
1885 if (!zalloc_cpumask_var(&osnoise_cpumask_new, GFP_KERNEL))
1888 err = cpulist_parse(buf, osnoise_cpumask_new);
1893 * trace_types_lock is taken to avoid concurrency on start/stop.
1895 mutex_lock(&trace_types_lock);
1896 running = osnoise_has_registered_instances();
1898 stop_per_cpu_kthreads();
1900 mutex_lock(&interface_lock);
1902 * osnoise_cpumask is read by CPU hotplug operations.
1906 cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new);
1909 mutex_unlock(&interface_lock);
1912 start_per_cpu_kthreads();
1913 mutex_unlock(&trace_types_lock);
1915 free_cpumask_var(osnoise_cpumask_new);
1919 free_cpumask_var(osnoise_cpumask_new);
1925 * osnoise/runtime_us: cannot be greater than the period.
1927 static struct trace_min_max_param osnoise_runtime = {
1928 .lock = &interface_lock,
1929 .val = &osnoise_data.sample_runtime,
1930 .max = &osnoise_data.sample_period,
1935 * osnoise/period_us: cannot be smaller than the runtime.
1937 static struct trace_min_max_param osnoise_period = {
1938 .lock = &interface_lock,
1939 .val = &osnoise_data.sample_period,
1941 .min = &osnoise_data.sample_runtime,
1945 * osnoise/stop_tracing_us: no limit.
1947 static struct trace_min_max_param osnoise_stop_tracing_in = {
1948 .lock = &interface_lock,
1949 .val = &osnoise_data.stop_tracing,
1955 * osnoise/stop_tracing_total_us: no limit.
1957 static struct trace_min_max_param osnoise_stop_tracing_total = {
1958 .lock = &interface_lock,
1959 .val = &osnoise_data.stop_tracing_total,
1964 #ifdef CONFIG_TIMERLAT_TRACER
1966 * osnoise/print_stack: print the stacktrace of the IRQ handler if the total
1967 * latency is higher than val.
1969 static struct trace_min_max_param osnoise_print_stack = {
1970 .lock = &interface_lock,
1971 .val = &osnoise_data.print_stack,
1977 * osnoise/timerlat_period: min 100 us, max 1 s
1979 u64 timerlat_min_period = 100;
1980 u64 timerlat_max_period = 1000000;
1981 static struct trace_min_max_param timerlat_period = {
1982 .lock = &interface_lock,
1983 .val = &osnoise_data.timerlat_period,
1984 .max = &timerlat_max_period,
1985 .min = &timerlat_min_period,
1989 static const struct file_operations cpus_fops = {
1990 .open = tracing_open_generic,
1991 .read = osnoise_cpus_read,
1992 .write = osnoise_cpus_write,
1993 .llseek = generic_file_llseek,
1996 #ifdef CONFIG_TIMERLAT_TRACER
1997 #ifdef CONFIG_STACKTRACE
1998 static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2002 tmp = tracefs_create_file("print_stack", TRACE_MODE_WRITE, top_dir,
2003 &osnoise_print_stack, &trace_min_max_fops);
2009 #else /* CONFIG_STACKTRACE */
2010 static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2014 #endif /* CONFIG_STACKTRACE */
2017 * init_timerlat_tracefs - A function to initialize the timerlat interface files
2019 static int init_timerlat_tracefs(struct dentry *top_dir)
2023 tmp = tracefs_create_file("timerlat_period_us", TRACE_MODE_WRITE, top_dir,
2024 &timerlat_period, &trace_min_max_fops);
2028 return init_timerlat_stack_tracefs(top_dir);
2030 #else /* CONFIG_TIMERLAT_TRACER */
2031 static int init_timerlat_tracefs(struct dentry *top_dir)
2035 #endif /* CONFIG_TIMERLAT_TRACER */
2038 * init_tracefs - A function to initialize the tracefs interface files
2040 * This function creates entries in tracefs for "osnoise" and "timerlat".
2041 * It creates these directories in the tracing directory, and within that
2042 * directory the use can change and view the configs.
2044 static int init_tracefs(void)
2046 struct dentry *top_dir;
2050 ret = tracing_init_dentry();
2054 top_dir = tracefs_create_dir("osnoise", NULL);
2058 tmp = tracefs_create_file("period_us", TRACE_MODE_WRITE, top_dir,
2059 &osnoise_period, &trace_min_max_fops);
2063 tmp = tracefs_create_file("runtime_us", TRACE_MODE_WRITE, top_dir,
2064 &osnoise_runtime, &trace_min_max_fops);
2068 tmp = tracefs_create_file("stop_tracing_us", TRACE_MODE_WRITE, top_dir,
2069 &osnoise_stop_tracing_in, &trace_min_max_fops);
2073 tmp = tracefs_create_file("stop_tracing_total_us", TRACE_MODE_WRITE, top_dir,
2074 &osnoise_stop_tracing_total, &trace_min_max_fops);
2078 tmp = trace_create_file("cpus", TRACE_MODE_WRITE, top_dir, NULL, &cpus_fops);
2082 ret = init_timerlat_tracefs(top_dir);
2089 tracefs_remove(top_dir);
2093 static int osnoise_hook_events(void)
2098 * Trace is already hooked, we are re-enabling from
2101 if (trace_osnoise_callback_enabled)
2104 retval = hook_irq_events();
2108 retval = hook_softirq_events();
2110 goto out_unhook_irq;
2112 retval = hook_thread_events();
2119 unhook_softirq_events();
2121 unhook_irq_events();
2126 * osnoise_workload_start - start the workload and hook to events
2128 static int osnoise_workload_start(void)
2133 * Instances need to be registered after calling workload
2134 * start. Hence, if there is already an instance, the
2135 * workload was already registered. Otherwise, this
2136 * code is on the way to register the first instance,
2137 * and the workload will start.
2139 if (osnoise_has_registered_instances())
2142 osn_var_reset_all();
2144 retval = osnoise_hook_events();
2149 * Make sure that ftrace_nmi_enter/exit() see reset values
2150 * before enabling trace_osnoise_callback_enabled.
2153 trace_osnoise_callback_enabled = true;
2155 retval = start_per_cpu_kthreads();
2157 unhook_irq_events();
2165 * osnoise_workload_stop - stop the workload and unhook the events
2167 static void osnoise_workload_stop(void)
2170 * Instances need to be unregistered before calling
2171 * stop. Hence, if there is a registered instance, more
2172 * than one instance is running, and the workload will not
2173 * yet stop. Otherwise, this code is on the way to disable
2174 * the last instance, and the workload can stop.
2176 if (osnoise_has_registered_instances())
2179 trace_osnoise_callback_enabled = false;
2181 * Make sure that ftrace_nmi_enter/exit() see
2182 * trace_osnoise_callback_enabled as false before continuing.
2186 stop_per_cpu_kthreads();
2188 unhook_irq_events();
2189 unhook_softirq_events();
2190 unhook_thread_events();
2193 static void osnoise_tracer_start(struct trace_array *tr)
2198 * If the instance is already registered, there is no need to
2199 * register it again.
2201 if (osnoise_instance_registered(tr))
2204 retval = osnoise_workload_start();
2206 pr_err(BANNER "Error starting osnoise tracer\n");
2208 osnoise_register_instance(tr);
2211 static void osnoise_tracer_stop(struct trace_array *tr)
2213 osnoise_unregister_instance(tr);
2214 osnoise_workload_stop();
2217 static int osnoise_tracer_init(struct trace_array *tr)
2220 * Only allow osnoise tracer if timerlat tracer is not running
2223 if (timerlat_enabled())
2226 tr->max_latency = 0;
2228 osnoise_tracer_start(tr);
2232 static void osnoise_tracer_reset(struct trace_array *tr)
2234 osnoise_tracer_stop(tr);
2237 static struct tracer osnoise_tracer __read_mostly = {
2239 .init = osnoise_tracer_init,
2240 .reset = osnoise_tracer_reset,
2241 .start = osnoise_tracer_start,
2242 .stop = osnoise_tracer_stop,
2243 .print_header = print_osnoise_headers,
2244 .allow_instances = true,
2247 #ifdef CONFIG_TIMERLAT_TRACER
2248 static void timerlat_tracer_start(struct trace_array *tr)
2253 * If the instance is already registered, there is no need to
2254 * register it again.
2256 if (osnoise_instance_registered(tr))
2259 retval = osnoise_workload_start();
2261 pr_err(BANNER "Error starting timerlat tracer\n");
2263 osnoise_register_instance(tr);
2268 static void timerlat_tracer_stop(struct trace_array *tr)
2272 osnoise_unregister_instance(tr);
2275 * Instruct the threads to stop only if this is the last instance.
2277 if (!osnoise_has_registered_instances()) {
2278 for_each_online_cpu(cpu)
2279 per_cpu(per_cpu_osnoise_var, cpu).sampling = 0;
2282 osnoise_workload_stop();
2285 static int timerlat_tracer_init(struct trace_array *tr)
2288 * Only allow timerlat tracer if osnoise tracer is not running already.
2290 if (osnoise_has_registered_instances() && !osnoise_data.timerlat_tracer)
2294 * If this is the first instance, set timerlat_tracer to block
2295 * osnoise tracer start.
2297 if (!osnoise_has_registered_instances())
2298 osnoise_data.timerlat_tracer = 1;
2300 tr->max_latency = 0;
2301 timerlat_tracer_start(tr);
2306 static void timerlat_tracer_reset(struct trace_array *tr)
2308 timerlat_tracer_stop(tr);
2311 * If this is the last instance, reset timerlat_tracer allowing
2312 * osnoise to be started.
2314 if (!osnoise_has_registered_instances())
2315 osnoise_data.timerlat_tracer = 0;
2318 static struct tracer timerlat_tracer __read_mostly = {
2320 .init = timerlat_tracer_init,
2321 .reset = timerlat_tracer_reset,
2322 .start = timerlat_tracer_start,
2323 .stop = timerlat_tracer_stop,
2324 .print_header = print_timerlat_headers,
2325 .allow_instances = true,
2328 __init static int init_timerlat_tracer(void)
2330 return register_tracer(&timerlat_tracer);
2332 #else /* CONFIG_TIMERLAT_TRACER */
2333 __init static int init_timerlat_tracer(void)
2337 #endif /* CONFIG_TIMERLAT_TRACER */
2339 __init static int init_osnoise_tracer(void)
2343 mutex_init(&interface_lock);
2345 cpumask_copy(&osnoise_cpumask, cpu_all_mask);
2347 ret = register_tracer(&osnoise_tracer);
2349 pr_err(BANNER "Error registering osnoise!\n");
2353 ret = init_timerlat_tracer();
2355 pr_err(BANNER "Error registering timerlat!\n");
2359 osnoise_init_hotplug_support();
2361 INIT_LIST_HEAD_RCU(&osnoise_instances);
2367 late_initcall(init_osnoise_tracer);